Automatic MRI Bone Segmentation Toki Migimatsu, Advisor: Gordon Wetzstein Department of Computer Science, Stanford University Motivation Segmentation Technique Challenges & Future Work Experimental Results Magnetic Resonance Imaging (MRI) provides a safe and non-invasive way to study internal tissues and create detailed musculoskeletal models of the body. Aside from clinical applications, these models can be used in areas ranging from character animation to assistive robotics, where accurate models of human motion are important. However, obtaining these musculoskeletal models requires manual segmentation, a prohibitively time-consuming process; segmenting one knee alone takes an expert hours [1] . Automatic segmentation could allow the mass generation of musculoskeletal models. Current methods for automatic segmentation rely on prior datasets to generate new segmentations [1]–[3] . However, large databases of manual segmentations are unavailable for most bones. Other methods are semi-automatic [4] [5] , and require rough manual segmentations to start. Examples of fully automatic segmentation do not exist in literature. Challenges in automatically segmenting MRI include low signal-to- noise, highly inconsistent lighting, and varying appearances of a single bone due to the trabecular and cortical tissues. Future work includes experimenting with an MSER algorithm that uses 3D level sets (instead of running 2D MSER along multiple axes), and refining the segmentation method to work on bones with smaller cortical layers. Mask: LoG, Morphological Close 3D Adaptive Equalization Trabecular: 2-pass Bright-on-Dark MSER Trabecular: Connected Components Cortical: Dark-on-Bright MSER Cortical: Connected Components Original Accuracy: 98.30% False positives: 0.60% False negatives: 0.10% Ground Truth Automatic Segmentation Error Accuracy: 97.78% False positives: 0.02% False negatives: 2.20% Ground Truth Error Automatic Segmentation Failed attempt: 3D Sobel [1] H. Seim, D. Kainmueller, H. Lamecker, M. Bindernagel, J. Malinowski, and S. Zachow, “Model-based auto-segmentation of knee bones and cartilage in mri data,” in Proc. MICCAI Workshop Medical Image Analysis for the Clinic , Ed., 2010. [2] J. Schmid and N. Magnenat-Thalmann, “Mri bone segmentation using deformable models and shape priors,” English, in Medical Image Computing and Computer-Assisted Intervention – MICCAI 2008, vol. 5241, Springer Berlin Heidelberg, 2008. [3] Y. Xia, S. Chandra, O. Salvado, J. Fripp, R. Schwarz, L. Lauer, C. Engstrom, and S. Crozier, “Automated mr hip bone segmentation,” in Digital Image Computing Techniques and Appli- cations (DICTA), 2011 International Conference on, Dec. 2011. [4] A. Rusu, Segmentation of bone structures in magnetic resonance images (mri) for human hand skeletal kinematics modelling, 2011. [5] E. Jolivet, E. Dion, P. Rouch, G. Dubois, R. Charrier, C. Payan, and W. Skalli, “Skeletal muscle segmentation from mri dataset using a model-based approach,” Computer Methods in Biomechanics and Biomedical Engineering: Imaging & Visualization, 2014.

Automatic MRI Bone Segmentation · robotics, where accurate models of human motion are important. However, obtaining these musculoskeletal models requires manual segmentation, a prohibitively

Download PDF Report

Upload
others
View
5
Download
0

Embed Size (px)

Citation preview

Page 1: Automatic MRI Bone Segmentation · robotics, where accurate models of human motion are important. However, obtaining these musculoskeletal models requires manual segmentation, a prohibitively

Automatic MRI Bone SegmentationToki Migimatsu, Advisor: Gordon WetzsteinDepartment of Computer Science, Stanford University

Motivation Segmentation Technique

Challenges & Future Work Experimental Results

Magnetic Resonance Imaging (MRI) provides a safe and non-invasive way to study internal tissues and create detailed musculoskeletal models of the body. Aside from clinical applications, these models can be used in areas ranging from character animation to assistive robotics, where accurate models of human motion are important.

However, obtaining these musculoskeletal models requires manual segmentation, a prohibitively time-consuming process; segmenting one knee alone takes an expert hours [1].

Automatic segmentation could allow the mass generation of musculoskeletal models.

Current methods for automatic segmentation rely on prior datasets to generate new segmentations [1]–[3]. However, large databases of manual segmentations are unavailable for most bones. Other methods are semi-automatic [4] [5], and require rough manual segmentations to start. Examples of fully automatic segmentation do not exist in literature.

Challenges in automatically segmenting MRI include low signal-to-noise, highly inconsistent lighting, and varying appearances of a single bone due to the trabecular and cortical tissues.

Future work includes experimenting with an MSER algorithm that uses 3D level sets (instead of running 2D MSER along multiple axes), and refining the segmentation method to work on bones with smaller cortical layers.

Mask: LoG, Morphological Close 3D Adaptive Equalization

Trabecular: 2-pass Bright-on-Dark MSERTrabecular: Connected ComponentsCortical: Dark-on-Bright MSERCortical: Connected Components

Original

Accuracy: 98.30%False positives: 0.60%False negatives: 0.10%

Ground Truth

Automatic Segmentation

Error

Accuracy: 97.78%False positives: 0.02%False negatives: 2.20%

Ground Truth Error

Automatic Segmentation

Failed attempt: 3D Sobel[1] H. Seim, D. Kainmueller, H. Lamecker, M. Bindernagel, J. Malinowski, and S. Zachow, “Model-based auto-segmentation of knee bones and cartilage in mri data,” in Proc. MICCAI Workshop Medical Image Analysis for the Clinic, Ed., 2010. [2] J. Schmid and N. Magnenat-Thalmann, “Mri bone segmentation using deformable models and shape priors,” English, in Medical Image Computing and Computer-Assisted Intervention – MICCAI 2008, vol. 5241, Springer Berlin Heidelberg, 2008. [3] Y. Xia, S. Chandra, O. Salvado, J. Fripp, R. Schwarz, L. Lauer, C. Engstrom, and S. Crozier, “Automated mr hip bone segmentation,” in Digital Image Computing Techniques and Appli- cations (DICTA), 2011 International Conference on, Dec. 2011. [4] A. Rusu, Segmentation of bone structures in magnetic resonance images (mri) for human hand skeletal kinematics modelling, 2011. [5] E. Jolivet, E. Dion, P. Rouch, G. Dubois, R. Charrier, C. Payan, and W. Skalli, “Skeletal muscle segmentation from mri dataset using a model-based approach,” Computer Methods in Biomechanics and Biomedical Engineering: Imaging & Visualization, 2014.